I've submitted 207175 for a clang 3.8.0 va_list handling problem for powerpc

[Mark Millard]

On 2016-Feb-15, at 12:18 PM, Roman Divacky <rdivacky at vlakno.cz> wrote:
> 
> On Mon, Feb 15, 2016 at 12:17:50PM -0800, Mark Millard wrote:
>> On 2016-Feb-15, at 11:11 AM, Roman Divacky <rdivacky at vlakno.cz> wrote:
>>> 
>>> Mark, I believe you're right. What do you think about this patch?
>>> 
>>> Index: tools/clang/lib/CodeGen/TargetInfo.cpp
>>> ===================================================================
>>> --- tools/clang/lib/CodeGen/TargetInfo.cpp	(revision 260852)
>>> +++ tools/clang/lib/CodeGen/TargetInfo.cpp	(working copy)
>>> @@ -3599,6 +3599,8 @@
>>>  {
>>>    CGF.EmitBlock(UsingOverflow);
>>> 
>>> +    Builder.CreateStore(Builder.getInt8(11), NumRegsAddr);
>>> +
>>>    // Everything in the overflow area is rounded up to a size of at least 4.
>>>    CharUnits OverflowAreaAlign = CharUnits::fromQuantity(4);
>>> 
>>> 
>>> Can you test it?
>> 
>> It may be later today before I can start the the test process.
>> 
>> While your change is not wrong as presented, it does seem to be based on the ABI document's numbering with the range 3 <= gr <12, where 3 <= gr < 11 cover r3-r10 use and gr=11 implies overflow stack area use. (gr being the ABI documents name.)
>> 
>> The clang code generation that I saw while analyzing the problem and the clang source that you had me look at did not use that numbering. Instead it seems to be based on 0 <= gpr < 9, where 0 <= gpr < 8 cover r3-r10 use and gpr=8 implies overflow stack area use. (gpr being what gdb showed me as I remember.) In other words: clang counts the number of "parameter registers" already in use as it goes along instead of tracking register numbers that have been used.
>> 
>> So assigning any value that appears to be positive and >= 8 should work, such as:
>> 
>> Builder.CreateStore(Builder.getInt8(8), NumRegsAddr);
> 
> Can you check what number gcc uses? We want to be interoperable with gcc.
> 
> Anyway, thanks for testing!
> 
> Roman

I'll do that check of gcc 4.2.1 code generation before starting the test later today.

But if the clang numbering is different in gcc 4.2.1 then far more than just adding a

> Builder.CreateStore(Builder.getInt8(?), NumRegsAddr)


for some "?" would need to be involved in the changes in order to reach compatibility.


I'll note that for clang 3.8.0 the actual comparison instruction generated is of the form

> cmplwi  r?,7


for some r?, such as r5 or r4, and the conditional branch generated is a bgt instruction.

===
Mark Millard
markmi at dsl-only.net